Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
Keep a wrapped Inductor pass's None uuid as a cache opt-out (#15429)
Daniel Hanbf183aab84ff
diff --git a/tests/test_inductor_readback_barrier.py b/tests/test_inductor_readback_barrier.pyindex 7fe2d0ee914..1a2b851ce1e 100644--- a/tests/test_inductor_readback_barrier.py+++ b/tests/test_inductor_readback_barrier.py@@ -358,3 +358,88 @@ def test_a_warm_cache_from_before_the_fix_is_not_served_with_a_user_pass(tmp_patpatched = run("patched")assert patched.get("fxgraph_cache_hit", 0) == 0assert patched.get("fxgraph_cache_miss", 0) >= 1+++def test_a_wrapped_user_pass_that_later_opts_out_of_caching_still_bypasses(monkeypatch):+ """torch's contract: uuid() -> None skips the FX cache. The wrapper must forward that."""+ from torch._inductor.custom_graph_pass import CustomGraphPass++ user = _user_pass_class()()+ _, inductor_config = _install_with(monkeypatch, user)+ keyed = inductor_config.post_grad_custom_post_pass+ assert keyed.uuid() == ("unsloth-inductor-readback-barrier-v1", "user-pass-v1")+ user.uuid = lambda: None+ assert isinstance(keyed, CustomGraphPass)+ assert keyed.uuid() is None+++_OPT_OUT_RUNNER = textwrap.dedent(+ """+ import importlib.util, json, sys, torch+ from torch._inductor import config+ from torch._inductor.custom_graph_pass import CustomGraphPass+ from torch._dynamo.utils import counters++ class TogglingPass(CustomGraphPass):+ enabled = True+ offset = 2.0++ def uuid(self):+ return "initial" if self.enabled else None++ def __call__(self, graph):+ for node in graph.nodes:+ if node.op == "call_function" and node.target == torch.ops.aten.add.Tensor:+ node.args = (node.args[0], self.offset)++ user = TogglingPass()+ config.post_grad_custom_post_pass = user+ if sys.argv[1] == "patched":+ spec = importlib.util.spec_from_file_location("fixes", sys.argv[2])+ fixes = importlib.util.module_from_spec(spec)+ spec.loader.exec_module(fixes)+ fixes._key_inductor_caches_on_readback_barrier(config, CustomGraphPass)+ user.enabled = False # opt out of caching, then change what the pass does+ results = []+ for offset in (2.0, 3.0):+ user.offset = offset+ torch._dynamo.reset()+ counters.clear()+ out = torch.compile(lambda x: x + 1, fullgraph = True)(torch.zeros(4))+ results.append({"out": out.tolist(), "inductor": dict(counters["inductor"])})+ print(json.dumps(results))+ """+)+++def test_a_cache_opted_out_user_pass_is_never_served_a_stale_graph(tmp_path):+ """End to end on CPU: after uuid() -> None, a changed pass must recompile, not hit the cache."""+ pytest.importorskip("torch._inductor.custom_graph_pass")+ import json++ runner = tmp_path / "runner.py"+ runner.write_text(_OPT_OUT_RUNNER)++ def run(arm):+ env = {+ **os.environ,+ "TORCHINDUCTOR_CACHE_DIR": str(tmp_path / f"inductor_{arm}"),+ "TORCHINDUCTOR_FX_GRAPH_CACHE": "1",+ "TORCHINDUCTOR_AUTOGRAD_CACHE": "1",+ "CUDA_VISIBLE_DEVICES": "",+ }+ out = subprocess.run(+ [sys.executable, str(runner), arm, _FIXES],+ env = env,+ capture_output = True,+ text = True,+ timeout = 600,+ )+ assert out.returncode == 0, out.stderr[-2000:]+ return json.loads(out.stdout.strip().splitlines()[-1])++ base = run("base")+ assert [r["out"] for r in base] == [[2.0] * 4, [3.0] * 4] # the harness can see the change+ patched = run("patched")+ assert [r["out"] for r in patched] == [[2.0] * 4, [3.0] * 4]+ assert all(r["inductor"].get("fxgraph_cache_hit", 0) == 0 for r in patched)diff --git a/unsloth/import_fixes.py b/unsloth/import_fixes.pyindex cf79fb6df48..a688a9373b0 100644--- a/unsloth/import_fixes.py+++ b/unsloth/import_fixes.py@@ -7106,7 +7106,9 @@ def _key_inductor_caches_on_readback_barrier(inductor_config, custom_graph_pass)return self.inner(graph, *args, **kwargs)def uuid(self):- return (marker, self.inner.uuid())+ # None is the inner pass opting out of the FX cache: keep it, never key a stale graph.+ inner_uuid = self.inner.uuid()+ return None if inner_uuid is None else (marker, inner_uuid)inductor_config.post_grad_custom_post_pass = ReadbackBarrierKeyedPass(existing)